diff --git a/classpath/java-lang.cpp b/classpath/java-lang.cpp
index 39098ca4c4..74a3c9d838 100644
--- a/classpath/java-lang.cpp
+++ b/classpath/java-lang.cpp
@@ -482,16 +482,3 @@ Java_java_lang_Double_fillBufferWithDouble(JNIEnv* e, jclass, jdouble val,
   e->ReleaseByteArrayElements(buffer, buf, 0);
   return count;
 }
-
-extern "C" JNIEXPORT void JNICALL
-Java_java_lang_Long_printLong(JNIEnv*, jclass, jlong v)
-{
-  fprintf(stderr, "DEBUG: %lld\n", v);
-}
-
-
-extern "C" JNIEXPORT void JNICALL
-Java_java_lang_Long_printInt(JNIEnv*, jclass, jint v)
-{
-  fprintf(stderr, "DEBUG: %d\n", v);
-}
diff --git a/classpath/java/lang/Long.java b/classpath/java/lang/Long.java
index 4c006ea549..013ac883bb 100644
--- a/classpath/java/lang/Long.java
+++ b/classpath/java/lang/Long.java
@@ -38,16 +38,11 @@ public final class Long extends Number implements Comparable<Long> {
     return String.valueOf(value);
   }
 
-  private static native void printLong(long v);
-  private static native void printInt(long v);
-
   public static String toString(long v, int radix) {
     if (radix < 1 || radix > 36) {
       throw new IllegalArgumentException("radix " + radix + " not in [1,36]");
     }
 
-    printLong(v);
-
     if (v == 0) {
       return "0";
     }
@@ -58,13 +53,10 @@ public final class Long extends Number implements Comparable<Long> {
     int size = (negative ? 1 : 0);
     for (long n = v; n > 0; n /= radix) ++size;
 
-    printInt(size);
-
     char[] array = new char[size];
 
     int i = array.length - 1;
     for (long n = v; n > 0; n /= radix) {
-      printLong(n);
       long digit = n % radix;
       if (digit >= 0 && digit <= 9) {
         array[i] = (char) ('0' + digit);
diff --git a/src/compile.cpp b/src/compile.cpp
index ee81de2aa6..5c9ad91a3f 100644
--- a/src/compile.cpp
+++ b/src/compile.cpp
@@ -18,7 +18,7 @@ vmJump(void* address, void* base, void* stack, void* thread);
 
 namespace {
 
-const bool Verbose = true;
+const bool Verbose = false;
 const bool DebugTraces = false;
 
 class MyThread: public Thread {
@@ -643,7 +643,7 @@ class Frame {
     assert(t, sp - 2 >= localSize(t, method));
     assert(t, getBit(map, sp - 1) == 0);
     assert(t, getBit(map, sp - 2) == 0);
-    return c->stack(stack, 1);
+    return c->stack(stack, 0);
   }
 
   Operand* topObject() {
@@ -695,8 +695,7 @@ class Frame {
     assert(t, index < codeMaxLocals(t, methodCode(t, method)));
     assert(t, index < parameterFootprint(t, method)
            or getBit(map, index - parameterFootprint(t, method)) == 0);
-    pushInt
-      (c->signExtend4(c->memory(c->base(), localOffset(t, index, method))));
+    pushInt(c->select4(c->memory(c->base(), localOffset(t, index, method))));
   }
 
   void loadLong(unsigned index) {
@@ -706,7 +705,11 @@ class Frame {
            or getBit(map, index - parameterFootprint(t, method)) == 0);
     assert(t, index < parameterFootprint(t, method)
            or getBit(map, index + 1 - parameterFootprint(t, method)) == 0);
-    pushLong(c->select8(c->memory(c->base(), localOffset(t, index, method))));
+    pushLong
+      (c->select8
+       (c->memory
+        (c->base(),
+         localOffset(t, index + (4 / BytesPerWord), method))));
   }
 
   void loadObject(unsigned index) {
@@ -722,7 +725,11 @@ class Frame {
   }
 
   void storeLong(unsigned index) {
-    popLong(c->select8(c->memory(c->base(), localOffset(t, index, method))));
+    popLong
+      (c->select8
+       (c->memory
+        (c->base(),
+         localOffset(t, index + (4 / BytesPerWord), method))));
     storedInt(index);
     storedInt(index + 1);
   }
@@ -1180,10 +1187,18 @@ compileThrowNew(MyThread* t, Frame* frame, Machine::Type type)
   frame->trace();
 }
 
+void
+checkCast(MyThread* t, object class_, object o)
+{
+  if (UNLIKELY(o and not isAssignableFrom(t, class_, objectClass(t, o)))) {
+    throwNew(t, arrayBody(t, t->m->types, Machine::ClassCastExceptionType));
+  }
+}
+
 void
 pushReturnValue(MyThread* t, Frame* frame, unsigned code)
 {
-  Object* result = c->result();
+  Operand* result = frame->c->result();
 
   switch (code) {
   case ByteField:
@@ -1192,7 +1207,7 @@ pushReturnValue(MyThread* t, Frame* frame, unsigned code)
   case ShortField:
   case FloatField:
   case IntField:
-    frame->pushInt(c->signExtend4(result));
+    frame->pushInt(frame->c->select4(result));
     break;
 
   case ObjectField:
@@ -1201,7 +1216,7 @@ pushReturnValue(MyThread* t, Frame* frame, unsigned code)
 
   case LongField:
   case DoubleField:
-    frame->pushLong(c->select8(result));
+    frame->pushLong(frame->c->select8(result));
     break;
 
   case VoidField:
@@ -1211,13 +1226,13 @@ pushReturnValue(MyThread* t, Frame* frame, unsigned code)
     abort(t);
   }
 
-  c->release(c);
+  frame->c->release(result);
 }
 
 void
 compileDirectInvoke(MyThread* t, Frame* frame, object target)
 {
-  Operand* result = frame->c->alignedCall
+  frame->c->alignedCall
     (frame->c->constant
      (reinterpret_cast<intptr_t>
       (&singletonBody(t, methodCompiled(t, target), 0))));
@@ -1226,7 +1241,7 @@ compileDirectInvoke(MyThread* t, Frame* frame, object target)
 
   frame->pop(methodParameterFootprint(t, target));
 
-  pushReturnValue(t, frame, methodReturnCode(t, target), result);
+  pushReturnValue(t, frame, methodReturnCode(t, target));
 }
 
 void
@@ -1280,15 +1295,15 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
       case faload:
       case iaload:
-        frame->pushInt(c->signExtend4(c->memory(array, ArrayBody, index, 4)));
+        frame->pushInt(c->select4(c->memory(array, ArrayBody, index, 4)));
         break;
 
       case baload:
-        frame->pushInt(c->signExtend1(c->memory(array, ArrayBody, index, 1)));
+        frame->pushInt(c->select1(c->memory(array, ArrayBody, index, 1)));
         break;
 
       case caload:
-        frame->pushInt(c->zeroExtend2(c->memory(array, ArrayBody, index, 2)));
+        frame->pushInt(c->select2z(c->memory(array, ArrayBody, index, 2)));
         break;
 
       case daload:
@@ -1297,7 +1312,7 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
         break;
 
       case saload:
-        frame->pushInt(c->signExtend2(c->memory(array, ArrayBody, index, 2)));
+        frame->pushInt(c->select2(c->memory(array, ArrayBody, index, 2)));
         break;
       }
 
@@ -1425,15 +1440,18 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
       c->mark(nonnegative);
 
-      Operand* r = c->indirectCall
+      c->indirectCall
          (c->constant(reinterpret_cast<intptr_t>(makeBlankObjectArray)),
           3, c->thread(), frame->append(class_), length);
 
+      Operand* result = c->result();
+
       c->release(length);
 
       frame->trace();
 
-      frame->pushObject(r);
+      frame->pushObject(result);
+      c->release(result);
     } break;
 
     case areturn: {
@@ -1489,68 +1507,63 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       object class_ = resolveClassInPool(t, codePool(t, code), index - 1);
       if (UNLIKELY(t->exception)) return;
 
-      Operand* next = c->label();
-
       Operand* instance = frame->topObject();
-      Operand* tmp = c->temporary();
-
-      c->mov(instance, tmp);
-
-      c->cmp(c->constant(0), tmp);
-      c->je(next);
 
       Operand* classOperand = frame->append(class_);
 
-      c->mov(c->memory(tmp), tmp);
-      c->and_(c->constant(PointerMask), tmp);
+      c->indirectCall
+         (c->constant(reinterpret_cast<intptr_t>(checkCast)),
+          3, c->thread(), classOperand, instance);
 
-      c->cmp(classOperand, tmp);
-      c->je(next);
-
-      Operand* result = c->directCall
-        (c->constant(reinterpret_cast<intptr_t>(isAssignableFrom)),
-         3, c->thread(), classOperand, tmp);
-
-      c->release(tmp);
-
-      c->cmp(c->constant(0), result);
-      c->jne(next);
-        
-      compileThrowNew(t, frame, Machine::ClassCastExceptionType);
-
-      c->mark(next);
+      frame->trace();
     } break;
 
     case d2f: {
       Operand* a = frame->popLong();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(doubleToFloat)), 1, a));
+
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(doubleToFloat)), 1, a);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case d2i: {
       Operand* a = frame->popLong();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(doubleToInt)), 1, a));
+
+      c->directCall
+         (c->constant(reinterpret_cast<intptr_t>(doubleToInt)), 1, a);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case d2l: {
       Operand* a = frame->popLong();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(doubleToLong)), 1, a));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(doubleToLong)), 1, a);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case dadd: {
       Operand* a = frame->popLong();
       Operand* b = frame->popLong();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(addDouble)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(addDouble)), 2, a, b);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1558,9 +1571,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case dcmpg: {
       Operand* a = frame->popLong();
       Operand* b = frame->popLong();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(compareDoublesG)), 2, a, b));
+      
+      c->directCall
+         (c->constant(reinterpret_cast<intptr_t>(compareDoublesG)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1568,9 +1585,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case dcmpl: {
       Operand* a = frame->popLong();
       Operand* b = frame->popLong();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(compareDoublesL)), 2, a, b));
+
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(compareDoublesL)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1586,9 +1607,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case ddiv: {
       Operand* a = frame->popLong();
       Operand* b = frame->popLong();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(divideDouble)), 2, a, b));
+      
+      c->directCall
+         (c->constant(reinterpret_cast<intptr_t>(divideDouble)), 2, a, b);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1596,27 +1621,39 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case dmul: {
       Operand* a = frame->popLong();
       Operand* b = frame->popLong();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(multiplyDouble)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(multiplyDouble)), 2, a, b);
+      
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
 
     case dneg: {
       Operand* a = frame->popLong();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(negateDouble)), 1, a));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(negateDouble)), 1, a);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case vm::drem: {
       Operand* a = frame->popLong();
       Operand* b = frame->popLong();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(moduloDouble)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(moduloDouble)), 2, a, b);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1624,9 +1661,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case dsub: {
       Operand* a = frame->popLong();
       Operand* b = frame->popLong();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(subtractDouble)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(subtractDouble)), 2, a, b);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1657,34 +1698,50 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
     case f2d: {
       Operand* a = frame->popInt();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(floatToDouble)), 1, a));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(floatToDouble)), 1, a);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case f2i: {
       Operand* a = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(floatToInt)), 1, a));
+      
+      c->directCall
+         (c->constant(reinterpret_cast<intptr_t>(floatToInt)), 1, a);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case f2l: {
       Operand* a = frame->popInt();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(floatToLong)), 1, a));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(floatToLong)), 1, a);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case fadd: {
       Operand* a = frame->popInt();
       Operand* b = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(addFloat)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(addFloat)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1692,9 +1749,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case fcmpg: {
       Operand* a = frame->popInt();
       Operand* b = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(compareFloatsG)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(compareFloatsG)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1702,9 +1763,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case fcmpl: {
       Operand* a = frame->popInt();
       Operand* b = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(compareFloatsL)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(compareFloatsL)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1724,9 +1789,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case fdiv: {
       Operand* a = frame->popInt();
       Operand* b = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(divideFloat)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(divideFloat)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1734,27 +1803,39 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case fmul: {
       Operand* a = frame->popInt();
       Operand* b = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(multiplyFloat)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(multiplyFloat)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
 
     case fneg: {
       Operand* a = frame->popLong();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(negateFloat)), 1, a));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(negateFloat)), 1, a);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case vm::frem: {
       Operand* a = frame->popInt();
       Operand* b = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(moduloFloat)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(moduloFloat)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1762,9 +1843,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     case fsub: {
       Operand* a = frame->popInt();
       Operand* b = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(subtractFloat)), 2, a, b));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(subtractFloat)), 2, a, b);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
       c->release(b);
     } break;
@@ -1793,23 +1878,23 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       case ByteField:
       case BooleanField:
         frame->pushInt
-          (c->signExtend1(c->memory(table, fieldOffset(t, field))));
+          (c->select1(c->memory(table, fieldOffset(t, field))));
         break;
 
       case CharField:
         frame->pushInt
-          (c->zeroExtend2(c->memory(table, fieldOffset(t, field))));
+          (c->select2z(c->memory(table, fieldOffset(t, field))));
         break;
 
       case ShortField:
         frame->pushInt
-          (c->signExtend2(c->memory(table, fieldOffset(t, field))));
+          (c->select2(c->memory(table, fieldOffset(t, field))));
         break;
 
       case FloatField:
       case IntField:
         frame->pushInt
-          (c->signExtend4(c->memory(table, fieldOffset(t, field))));
+          (c->select4(c->memory(table, fieldOffset(t, field))));
         break;
 
       case DoubleField:
@@ -1849,39 +1934,47 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
     case i2b: {
       Operand* top = frame->topInt();
-      c->mov(c->signExtend1(top), top);
+      c->mov(c->select1(top), top);
     } break;
 
     case i2c: {
       Operand* top = frame->topInt();
-      c->mov(c->zeroExtend2(top), top);
+      c->mov(c->select2z(top), top);
     } break;
 
     case i2d: {
       Operand* a = frame->popInt();
-      frame->pushLong
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(intToDouble)), 1, a));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(intToDouble)), 1, a);
+
+      Operand* result = c->select8(c->result());
+      frame->pushLong(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case i2f: {
       Operand* a = frame->popInt();
-      frame->pushInt
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(intToFloat)), 1, a));
+      
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(intToFloat)), 1, a);
+
+      Operand* result = c->select4(c->result());
+      frame->pushInt(result);
+      c->release(result);
       c->release(a);
     } break;
 
     case i2l: {
       Operand* a = frame->popInt();
-      frame->pushLong(c->signExtend4(a));
+      frame->pushLong(c->select8From4(a));
       c->release(a);
     } break;
 
     case i2s: {
       Operand* top = frame->topInt();
-      c->mov(c->signExtend2(top), top);
+      c->mov(c->select2(top), top);
     } break;
       
     case iadd: {
@@ -2096,50 +2189,18 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       object class_ = resolveClassInPool(t, codePool(t, code), index - 1);
       if (UNLIKELY(t->exception)) return;
 
-      Operand* call = c->label();
-      Operand* next = c->label();
-      Operand* zero = c->label();
-
       Operand* instance = frame->popObject();
-      Operand* tmp = c->temporary();
-      Operand* result = c->temporary();
-
-      c->mov(instance, tmp);
-
-      c->cmp(c->constant(0), tmp);
-      c->je(zero);
 
       Operand* classOperand = frame->append(class_);
 
-      c->mov(c->memory(tmp), tmp);
-      c->and_(c->constant(PointerMask), tmp);
+      c->directCall
+         (c->constant(reinterpret_cast<intptr_t>(instanceOf)),
+          3, c->thread(), classOperand, instance);
 
-      c->cmp(classOperand, tmp);
-      c->jne(call);
-
-      c->mov(c->constant(1), result);
-      c->jmp(next);
-
-      c->mark(call);
-
-      c->mov
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(isAssignableFrom)),
-          3, c->thread(), classOperand, tmp), result);
-
-      c->release(tmp);
-      c->release(instance);
-
-      c->jmp(next);
-        
-      c->mark(zero);
-
-      c->mov(c->constant(0), result);
-
-      c->mark(next);
+      Operand* result = c->select4(c->result());
       frame->pushInt(result);
-
       c->release(result);
+      c->release(instance);
     } break;
 
     case invokeinterface: {
@@ -2153,18 +2214,21 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
       unsigned instance = parameterFootprint - 1;
 
-      Operand* result = c->call
-        (c->directCall
+      c->directCall
          (c->constant
           (reinterpret_cast<intptr_t>(findInterfaceMethodFromInstance)),
           3, c->thread(), frame->append(target),
-          c->stack(frame->stack, instance)));
+          c->stack(frame->stack, instance));
+
+      Operand* result = c->result();
+      c->call(result);
+      c->release(result);
       
       frame->trace(target, true);
 
       frame->pop(parameterFootprint);
 
-      pushReturnValue(t, frame, methodReturnCode(t, target), result);
+      pushReturnValue(t, frame, methodReturnCode(t, target));
     } break;
 
     case invokespecial: {
@@ -2213,7 +2277,7 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       c->mov(c->memory(instance), class_);
       c->and_(c->constant(PointerMask), class_);
 
-      Operand* result = c->call(c->memory(class_, offset));
+      c->call(c->memory(class_, offset));
 
       frame->trace(target, true);
 
@@ -2221,7 +2285,7 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
       frame->pop(parameterFootprint);
 
-      pushReturnValue(t, frame, methodReturnCode(t, target), result);
+      pushReturnValue(t, frame, methodReturnCode(t, target));
     } break;
 
     case ior: {
@@ -2349,11 +2413,11 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
     } break;
 
     case lconst_0:
-      frame->pushLong(c->constant(0));
+      frame->pushLong(c->select8(c->constant(0)));
       break;
 
     case lconst_1:
-      frame->pushLong(c->constant(1));
+      frame->pushLong(c->select8(c->constant(1)));
       break;
 
     case ldc:
@@ -2392,7 +2456,7 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
       uint64_t v;
       memcpy(&v, &singletonValue(t, pool, index - 1), 8);
-      frame->pushLong(c->constant(v));
+      frame->pushLong(c->select8(c->constant(v)));
     } break;
 
     case ldiv_: {
@@ -2469,10 +2533,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       }
       assert(t, start);
 
-      c->jmp
-        (c->directCall
-         (c->constant(reinterpret_cast<intptr_t>(lookUpAddress)),
-          4, key, start, c->constant(pairCount), default_));
+      c->directCall
+        (c->constant(reinterpret_cast<intptr_t>(lookUpAddress)),
+         4, key, start, c->constant(pairCount), default_);
+
+      Operand* result = c->result();
+      c->jmp(result);
+      c->release(result);
 
       c->release(key);
 
@@ -2587,15 +2654,18 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       if (UNLIKELY(t->exception)) return;
       PROTECT(t, class_);
 
-      Operand* result = c->indirectCall
+      c->indirectCall
         (c->constant(reinterpret_cast<intptr_t>(makeMultidimensionalArray)),
          4, c->thread(), frame->append(class_), c->stack(),
          c->constant(dimensions));
 
+      Operand* result = c->result();
+
       frame->trace();
 
       frame->pop(dimensions);
       frame->pushObject(result);
+      c->release(result);
     } break;
 
     case new_: {
@@ -2608,20 +2678,22 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       initClass(t, class_);
       if (UNLIKELY(t->exception)) return;
 
-      Operand* result;
       if (classVmFlags(t, class_) & WeakReferenceFlag) {
-        result = c->indirectCall
+        c->indirectCall
           (c->constant(reinterpret_cast<intptr_t>(makeNewWeakReference)),
            2, c->thread(), frame->append(class_));
       } else {
-        result = c->indirectCall
+        c->indirectCall
           (c->constant(reinterpret_cast<intptr_t>(makeNew)),
            2, c->thread(), frame->append(class_));
       }
 
+      Operand* result = c->result();
+
       frame->trace();
 
       frame->pushObject(result);
+      c->release(result);
     } break;
 
     case newarray: {
@@ -2675,16 +2747,19 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       default: abort(t);
       }
 
-      Operand* result = c->indirectCall
+      c->indirectCall
         (c->constant(reinterpret_cast<intptr_t>(makeBlankArray)),
          3, c->thread(), c->constant(reinterpret_cast<intptr_t>(constructor)),
          size);
+      
+      Operand* result = c->result();
 
       c->release(size);
 
       frame->trace();
 
       frame->pushObject(result);
+      c->release(result);
     } break;
 
     case nop: break;
@@ -3007,13 +3082,13 @@ finish(MyThread* t, Compiler* c, object method, Vector* objectPool,
     }
 
     // for debugging:
-    if (//false and
+    if (false and
         strcmp(reinterpret_cast<const char*>
                (&byteArrayBody(t, className(t, methodClass(t, method)), 0)),
-               "java/lang/Long") == 0 and
+               "java/lang/String") == 0 and
         strcmp(reinterpret_cast<const char*>
                (&byteArrayBody(t, methodName(t, method), 0)),
-               "toString") == 0)
+               "getBytes") == 0)
     {
       asm("int3");
     }
@@ -3182,7 +3257,11 @@ invokeNative2(MyThread* t, object method)
 
     case INT64_TYPE:
     case DOUBLE_TYPE: {
-      memcpy(args + argOffset, sp, 8);
+      if (BytesPerWord == 8) {
+        memcpy(args + argOffset, sp, 8);
+      } else {
+        memcpy(args + argOffset, sp - 1, 8);
+      }
       argOffset += (8 / BytesPerWord);
       sp -= 2;
     } break;
@@ -3363,10 +3442,13 @@ compileDefault(MyThread* t, Compiler* c)
   c->mov(c->base(), c->memory(c->thread(), difference(&(t->base), t)));
   c->mov(c->stack(), c->memory(c->thread(), difference(&(t->stack), t)));
 
-  c->jmp
-    (c->directCall
+  c->directCall
      (c->constant(reinterpret_cast<intptr_t>(compileMethod)),
-      1, c->thread()));
+      1, c->thread());
+
+  Operand* result = c->result();
+  c->jmp(result);
+  c->release(result);
 
   return finish(t, c);
 }
diff --git a/src/compiler.cpp b/src/compiler.cpp
index c622a687cf..31f4b06c8b 100644
--- a/src/compiler.cpp
+++ b/src/compiler.cpp
@@ -29,12 +29,10 @@ enum Register {
 enum SelectionType {
   Select1,
   Select2,
+  Select2z,
   Select4,
   Select8,
-  SignExtend1,
-  SignExtend2,
-  ZeroExtend2,
-  SignExtend4
+  Select8From4
 };
 
 const bool Verbose = false;
@@ -42,7 +40,7 @@ const bool Verbose = false;
 const unsigned RegisterCount = BytesPerWord * 2;
 const unsigned GprParameterCount = 6;
 const SelectionType DefaultSelection
-= (BytesPerWord == 8 ? S8Selection : S4Selection);
+= (BytesPerWord == 8 ? Select8 : Select4);
 
 class Context;
 class MyOperand;
@@ -55,6 +53,19 @@ class CodePromise;
 class MyPromise;
 class RegisterReference;
 
+inline unsigned
+footprint(SelectionType selection)
+{
+  switch (selection) {
+  case Select8:
+  case Select8From4:
+    return 8;
+
+  default:
+    return BytesPerWord;
+  }
+}
+
 int64_t
 divideLong(int64_t a, int64_t b)
 {
@@ -320,7 +331,7 @@ RegisterOperand*
 register_(Context* c, RegisterReference*, SelectionType = DefaultSelection);
 
 RegisterOperand*
-register_(Context* c, Register, Register = NoRegister,
+register_(Context* c, Register = NoRegister, Register = NoRegister,
           SelectionType = DefaultSelection);
 
 MemoryOperand*
@@ -429,38 +440,63 @@ release(Context* c, Register v)
 class RegisterReference {
  public:
   RegisterReference(Register value = NoRegister, Register high = NoRegister):
-    value_(value), high_(high), acquireHigh(false)
+    value_(value), defaultValue(value), high_(high), defaultHigh(high),
+    acquired(true)
   { }
 
   void acquire(Context* c) {
-    value_ = ::acquire(c);
-    if (acquireHigh) {
-      high_ = ::acquire(c);
+    if (defaultValue != NoRegister) {
+      ::acquire(c, defaultValue);
     }
+
+    if (defaultHigh != NoRegister) {
+      ::acquire(c, defaultHigh);
+    }
+
+    value_ = defaultValue;
+    high_ = defaultHigh;
+
+    acquired = true;
   }
 
   void release(Context* c) {
-    ::release(c, value_);
-    value_ = NoRegister;
+    assert(c, acquired);
+
+    if (value_ != NoRegister) {
+      ::release(c, value_);
+    }
 
     if (high_ != NoRegister) {
       ::release(c, high_);
     }
+
+    value_ = NoRegister;
+    high_ = NoRegister;
+
+    acquired = false;
   }
 
-  Register value(Context* c UNUSED) {
-    assert(c, value_ != NoRegister);
+  Register value(Context* c) {
+    assert(c, acquired);
+    if (value_ == NoRegister) {
+      value_ = ::acquire(c);
+    }
     return value_;
   }
 
-  Register high(Context* c UNUSED) {
-    assert(c, high_ != NoRegister);
+  Register high(Context* c) {
+    assert(c, acquired);
+    if (high_ == NoRegister) {
+      high_ = ::acquire(c);
+    }
     return high_;
   }
 
   Register value_;
+  Register defaultValue;
   Register high_;
-  bool acquireHigh;
+  Register defaultHigh;
+  bool acquired;
 };
 
 class RegisterOperand: public MyOperand {
@@ -478,7 +514,7 @@ class RegisterOperand: public MyOperand {
   }
 
   virtual unsigned footprint(Context*) {
-    return (selection == S8Selection ? 8 : BytesPerWord);
+    return ::footprint(selection);
   }
 
   virtual Register asRegister(Context* c) {
@@ -489,9 +525,6 @@ class RegisterOperand: public MyOperand {
     if (selection == this->selection) {
       return this;
     } else {
-      if (selection == S8Selection and BytesPerWord == 4) {
-        reference->acquireHigh = true;
-      }
       return register_(c, reference, selection);
     }
   }
@@ -528,7 +561,7 @@ class ImmediateOperand: public MyOperand {
   { }
 
   virtual unsigned footprint(Context*) {
-    return (selection == S8Selection ? 8 : BytesPerWord);
+    return ::footprint(selection);
   }
 
   virtual MyOperand* select(Context* c, SelectionType selection) {
@@ -610,7 +643,7 @@ class MemoryOperand: public MyOperand {
   { }
 
   virtual unsigned footprint(Context*) {
-    return (selection == S8Selection ? 8 : BytesPerWord);
+    return ::footprint(selection);
   }
 
   virtual Register asRegister(Context*);
@@ -698,7 +731,7 @@ temporary(Context* c)
 RegisterOperand*
 temporary(Context* c, SelectionType selection)
 {
-  if (BytesPerWord == 4 and selection == S8Selection) {
+  if (BytesPerWord == 4 and selection == Select8) {
     return register_(c, acquire(c), acquire(c), selection);
   } else {
     return register_(c, acquire(c), NoRegister, selection);
@@ -712,6 +745,14 @@ temporary(Context* c, Register v)
   return register_(c, v);
 }
 
+RegisterOperand*
+temporary(Context* c, Register v, Register h)
+{
+  acquire(c, v);
+  acquire(c, h);
+  return register_(c, v, h);
+}
+
 Segment*
 currentSegment(Context* c)
 {
@@ -793,6 +834,9 @@ class AcquireEvent: public Event {
   { }
 
   virtual void run(Context* c) {
+    if (Verbose) {
+      fprintf(stderr, "acquire register\n");
+    }
     operand->reference->acquire(c);
   }
 
@@ -807,6 +851,9 @@ class ReleaseEvent: public Event {
   { }
 
   virtual void run(Context* c) {
+    if (Verbose) {
+      fprintf(stderr, "release register\n");
+    }
     operand->release(c);
   }
 
@@ -954,18 +1001,38 @@ appendArgumentEvent(Context* c, MyOperand** arguments, unsigned count)
     ArgumentEvent(arguments, count, s->event);
 }
 
-MyStack*
-pushed(Context* c, MyStack* stack, unsigned footprint)
+void
+logStack(Context* c, MyStack* stack)
 {
-  int index = (stack ?
-               stack->index + (stack->value->footprint(c) / BytesPerWord) :
-               0);
+  fprintf(stderr, "ip %3d: ", currentSegment(c)->logicalIp);
+
+  if (stack) {
+    fprintf(stderr, " %d",
+            static_cast<MemoryOperand*>(stack->value)->displacement);
+  }
+
+  for (MyStack* s = stack; s; s = s->next) {
+    fprintf(stderr, "*");
+  }
+
+  fprintf(stderr, "\n");
+}
+
+MyStack*
+pushed(Context* c, MyStack* stack)
+{
+  int index = (stack ? stack->index + 1 : 0);
 
   MyOperand* value = memory
-    (c, register_(c, rbp), - (c->reserved + index + 1) * BytesPerWord, 0, 1,
-     footprint == (BytesPerWord * 2) ? S8Selection : DefaultSelection);
+    (c, register_(c, rbp), - (c->reserved + index + 1) * BytesPerWord, 0, 1);
 
-  return new (c->zone.allocate(sizeof(MyStack))) MyStack(value, index, stack);
+  stack = new (c->zone.allocate(sizeof(MyStack))) MyStack(value, index, stack);
+
+  if (Verbose) {
+    logStack(c, stack);
+  }
+
+  return stack;
 }
 
 MyStack*
@@ -973,7 +1040,10 @@ push(Context* c, MyStack* stack, MyOperand* v)
 {
   appendOperation(c, MyOperand::push, v);
 
-  return pushed(c, stack, v->footprint(c));
+  if ((v->footprint(c) / BytesPerWord) == 2) {
+    stack = pushed(c, stack);
+  }
+  return pushed(c, stack);
 }
 
 MyStack*
@@ -983,11 +1053,15 @@ pop(Context* c, MyStack* stack, int count)
     (c, MyOperand::add, immediate(c, count * BytesPerWord), register_(c, rsp));
 
   while (count) {
-    count -= (stack->value->footprint(c) / BytesPerWord);
+    -- count;
     assert(c, count >= 0);
     stack = stack->next;
   }
 
+  if (Verbose) {
+    logStack(c, stack);
+  }
+
   return stack;
 }
 
@@ -996,6 +1070,14 @@ pop(Context* c, MyStack* stack, MyOperand* dst)
 {
   appendOperation(c, MyOperand::pop, dst);
 
+  if (dst->footprint(c) / BytesPerWord == 2) {
+    stack = stack->next;
+  }
+
+  if (Verbose) {
+    logStack(c, stack->next);
+  }
+
   return stack->next;
 }
 
@@ -1031,6 +1113,17 @@ rex(Context* c)
   }
 }
 
+void
+rex(Context* c, SelectionType selection)
+{
+  if (selection == Select8) {
+    assert(c, BytesPerWord == 8);
+    rex(c);
+  } else {
+    assert(c, selection == Select4);
+  }
+}
+
 void
 encode(Context* c, uint8_t* instruction, unsigned length, int a, Register b,
        int32_t displacement, int index, unsigned scale)
@@ -1077,6 +1170,15 @@ encode(Context* c, uint8_t instruction, int a, MemoryOperand* b, bool rex)
   encode(c, &instruction, 1, a, r, b->displacement, index, b->scale);
 }
 
+void
+encode(Context* c, uint8_t instruction, int a, MemoryOperand* b,
+       SelectionType selection)
+{
+  assert(c, selection == DefaultSelection or selection == Select4);
+
+  encode(c, instruction, a, b, selection == DefaultSelection);
+}
+
 void
 encode2(Context* c, uint16_t instruction, int a, MemoryOperand* b, bool rex)
 {
@@ -1092,51 +1194,64 @@ encode2(Context* c, uint16_t instruction, int a, MemoryOperand* b, bool rex)
 void
 RegisterOperand::apply(Context* c, Operation operation)
 {
-  assert(c, operation == push
-         or operation == pop
-         or selection == DefaultSelection);
-
   switch (operation) {
   case call:
+    assert(c, selection == DefaultSelection);
+
     c->code.append(0xff);
     c->code.append(0xd0 | value(c));
     break;
 
   case jmp:
+    assert(c, selection == DefaultSelection);
+
     c->code.append(0xff);
     c->code.append(0xe0 | value(c));
     break;
 
   case pop:
-    if (selection == DefaultSelection) {
+    if (selection == DefaultSelection or selection == Select4) {
       c->code.append(0x58 | value(c));
     } else {
-      switch (selection) {
-      case S8Selection:
-        assert(c, selection == S8Selection);
+      assert(c, selection == Select8);
 
-        register_(c, value(c))->apply(c, pop);
-        register_(c, high(c))->apply(c, pop);
+      register_(c, value(c))->apply(c, pop);
+      register_(c, high(c))->apply(c, pop);
+    }
+    break;
+
+  case push:
+    if (selection == DefaultSelection or selection == Select4
+        or (BytesPerWord == 8 and selection == Select8From4))
+    {
+      c->code.append(0x50 | value(c));
+    } else {
+      switch (selection) {
+      case Select8:
+        register_(c, high(c))->apply(c, push);
+        register_(c, value(c))->apply(c, push);
         break;
 
+      case Select8From4: {
+        RegisterOperand* ax = temporary(c, rax);
+        RegisterOperand* dx = temporary(c, rdx);
+        
+        ax->accept(c, mov, register_(c, value(c)));
+        c->code.append(0x99); // cdq
+        dx->apply(c, push);
+        ax->apply(c, push);
+
+        ax->release(c);
+        dx->release(c);
+      } break;
+
       default: abort(c);
       }
     }
     break;
 
-  case push:
-    if (selection == DefaultSelection or selection == S4Selection) {
-      c->code.append(0x50 | value(c));
-    } else {
-      assert(c, selection == S8Selection);
-
-      register_(c, high(c))->apply(c, push);
-      register_(c, value(c))->apply(c, push);
-    }
-    break;
-
   case neg:
-    rex(c);
+    rex(c, selection);
     c->code.append(0xf7);
     c->code.append(0xd8 | value(c));
     break;
@@ -1151,21 +1266,27 @@ RegisterOperand::accept(Context* c, Operation operation,
 {
   switch (operation) {
   case add:
-    rex(c);
+    assert(c, selection == operand->selection);
+  
+    rex(c, operand->selection);
     c->code.append(0x01);
     c->code.append(0xc0 | (operand->value(c) << 3) | value(c));
     break;
 
   case cmp:
-    if (selection == DefaultSelection) {
-      rex(c);
+    assert(c, selection == operand->selection);
+  
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      rex(c, operand->selection);
       c->code.append(0x39);
       c->code.append(0xc0 | (operand->value(c) << 3) | value(c));
     } else {
-      assert(c, selection == S8Selection);
+      assert(c, operand->selection == Select8);
 
       register_(c, high(c))->accept
-        (c, mov, register_(c, operand->high(c)));
+        (c, cmp, register_(c, operand->high(c)));
 
       // if the high order bits are equal, we compare the low order
       // bits; otherwise, we jump past that comparison
@@ -1174,41 +1295,41 @@ RegisterOperand::accept(Context* c, Operation operation,
       c->code.append4(2);
 
       register_(c, value(c))->accept
-        (c, mov, register_(c, operand->value(c)));
+        (c, cmp, register_(c, operand->value(c)));
     }
     break;
 
   case mov:
     if (value(c) != operand->value(c) or selection != operand->selection) {
-      if (operand->selection == DefaultSelection) {
-        rex(c);
+      if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+      {
+        rex(c, selection);
         c->code.append(0x89);
         c->code.append(0xc0 | (operand->value(c) << 3) | value(c));
       } else {
         switch (operand->selection) {
-        case S1Selection:
+        case Select1:
           c->code.append(0xbe);
           c->code.append(0xc0 | (operand->value(c) << 3) | value(c));
           break;
 
-        case S2Selection:
+        case Select2:
           c->code.append(0xbf);
           c->code.append(0xc0 | (operand->value(c) << 3) | value(c));
           break;
 
-        case Z2Selection:
+        case Select2z:
           c->code.append(0xb7);
           c->code.append(0xc0 | (operand->value(c) << 3) | value(c));
           break;
 
-        case S4Selection:
+        case Select4:
           c->code.append(0x89);
           c->code.append(0xc0 | (operand->value(c) << 3) | value(c));
           break;
 
-        case S8Selection:
-          assert(c, selection == S8Selection);
-          
+        case Select8:
           register_(c, value(c))->accept
             (c, mov, register_(c, operand->value(c)));
 
@@ -1223,7 +1344,9 @@ RegisterOperand::accept(Context* c, Operation operation,
     break;
 
   case mul:
-    rex(c);
+    assert(c, selection == operand->selection);
+  
+    rex(c, selection);
     c->code.append(0x0f);
     c->code.append(0xaf);
     c->code.append(0xc0 | (value(c) << 3) | operand->value(c));
@@ -1349,6 +1472,8 @@ void
 RegisterOperand::accept(Context* c, Operation operation,
                         AddressOperand* operand)
 {
+  assert(c, selection == DefaultSelection);
+
   switch (operation) {
   case mov: {
     accept(c, operation, ::value(c, operand));
@@ -1362,38 +1487,43 @@ void
 RegisterOperand::accept(Context* c, Operation operation,
                         MemoryOperand* operand)
 {
-  assert(c, operation == mov or selection == DefaultSelection);
-  assert(c, operation == mov or operand->selection == DefaultSelection);
+  assert(c, selection == DefaultSelection);
 
   switch (operation) {
   case cmp: {
-    encode(c, 0x3b, value(c), operand, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {    
+      encode(c, 0x3b, value(c), operand, operand->selection);
+    } else {
+      abort(c);
+    }
   } break;
 
   case mov: {
-    if (operand->selection == DefaultSelection) {
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
       encode(c, 0x8b, value(c), operand, true);
     } else {
       switch (operand->selection) {
-      case S1Selection:
+      case Select1:
         encode2(c, 0x0fbe, value(c), operand, true);      
         break;
 
-      case S2Selection:
+      case Select2:
         encode2(c, 0x0fbf, value(c), operand, true);      
         break;
 
-      case Z2Selection:
+      case Select2z:
         encode2(c, 0x0fb7, value(c), operand, true);      
         break;
 
-      case S4Selection:
+      case Select4:
         encode(c, 0x63, value(c), operand, true);
         break;
 
-      case S8Selection:
-        assert(c, selection == S8Selection);
-
+      case Select8:
         register_(c, value(c))->accept
           (c, mov, memory
            (c, operand->base, operand->displacement,
@@ -1551,12 +1681,12 @@ AddressOperand::asRegister(Context* c)
 void
 ImmediateOperand::apply(Context* c, Operation operation)
 {
-  assert(c, operation == push or selection == DefaultSelection);
-
   switch (operation) {
   case alignedCall:
   case call:
   case jmp:
+    assert(c, selection == DefaultSelection);
+
     address(c, new (c->zone.allocate(sizeof(ResolvedPromise)))
             ResolvedPromise(value))->apply(c, operation);
     break;
@@ -1576,7 +1706,7 @@ ImmediateOperand::apply(Context* c, Operation operation)
         tmp->release(c);
       }
     } else {
-      assert(c, selection == S8Selection);
+      assert(c, selection == Select8);
 
       immediate(c, (value >> 32) & 0xFFFFFFFF)->apply(c, push);
       immediate(c, (value      ) & 0xFFFFFFFF)->apply(c, push);
@@ -1634,18 +1764,22 @@ MemoryOperand::apply(Context* c, Operation operation)
 {
   switch (operation) {
   case call:
+    assert(c, selection == DefaultSelection);
+
     encode(c, 0xff, 2, this, false);
     break;
 
   case jmp:
+    assert(c, selection == DefaultSelection);
+
     encode(c, 0xff, 4, this, false);
     break;
 
   case neg:
     if (selection == DefaultSelection) {
-      encode(c, 0xf7, 2, this, true);
+      encode(c, 0xf7, 2, this, selection);
     } else {
-      assert(c, selection == S8Selection);
+      assert(c, selection == Select8);
 
       RegisterOperand* ax = temporary(c, rax);
       RegisterOperand* dx = temporary(c, rdx);
@@ -1670,7 +1804,18 @@ MemoryOperand::apply(Context* c, Operation operation)
     break;
 
   case pop:
-    encode(c, 0x8f, 0, this, false);
+    if (selection == DefaultSelection) {
+      encode(c, 0x8f, 0, this, false);
+    } else {
+      assert(c, selection == Select8);
+
+      MemoryOperand* low = memory(c, base, displacement, index, scale);
+      MemoryOperand* high = memory
+        (c, base, displacement + BytesPerWord, index, scale);
+
+      low->apply(c, pop);
+      high->apply(c, pop);
+    }
     break;
 
   case push:
@@ -1678,7 +1823,7 @@ MemoryOperand::apply(Context* c, Operation operation)
       encode(c, 0xff, 6, this, false);
     } else {
       switch (selection) {
-      case S8Selection: {
+      case Select8: {
         MemoryOperand* low = memory(c, base, displacement, index, scale);
         MemoryOperand* high = memory
           (c, base, displacement + BytesPerWord, index, scale);
@@ -1689,7 +1834,7 @@ MemoryOperand::apply(Context* c, Operation operation)
         
       default: {
         RegisterOperand* tmp = temporary
-          (c, selection == S8Selection ? S8Selection : DefaultSelection);
+          (c, selection == Select8 ? Select8 : DefaultSelection);
         tmp->accept(c, mov, this);
         tmp->apply(c, operation);
         tmp->release(c);
@@ -1708,14 +1853,22 @@ MemoryOperand::accept(Context* c, Operation operation,
 {
   switch (operation) {
   case and_: {
-    encode(c, 0x21, operand->value(c), this, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      encode(c, 0x21, operand->value(c), this, operand->selection);
+    } else {
+      abort(c);
+    }
   } break;
 
   case add: {
-    if (selection == DefaultSelection) {
-      encode(c, 0x01, operand->value(c), this, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      encode(c, 0x01, operand->value(c), this, operand->selection);
     } else {
-      assert(c, selection == S8Selection);
+      assert(c, operand->selection == Select8);
 
       RegisterOperand* ax = temporary(c, rax);
       RegisterOperand* dx = temporary(c, rdx);
@@ -1733,22 +1886,26 @@ MemoryOperand::accept(Context* c, Operation operation,
   } break;
 
   case addc:
-    if (operand->selection == DefaultSelection) {
-      encode(c, 0x11, operand->value(c), this, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      encode(c, 0x11, operand->value(c), this, operand->selection);
     } else {
       abort(c);
     }
     break;
 
   case div: {
-    if (selection == DefaultSelection) {
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
       RegisterOperand* ax = temporary(c, rax);
       RegisterOperand* dx = temporary(c, rdx);
       ax->accept(c, mov, this);
     
-      rex(c);
+      rex(c, operand->selection);
       c->code.append(0x99);
-      rex(c);
+      rex(c, operand->selection);
       c->code.append(0xf7);
       c->code.append(0xf8 | operand->value(c));
 
@@ -1757,40 +1914,52 @@ MemoryOperand::accept(Context* c, Operation operation,
       ax->release(c);
       dx->release(c);
     } else {
-      assert(c, selection == S8Selection);
+      assert(c, operand->selection == Select8);
 
       operand->apply(c, push);
-      apply(c, push);
+      memory(c, base, displacement, index, scale, Select8)->apply(c, push);
       immediate(c, reinterpret_cast<intptr_t>(divideLong))->apply(c, call);
-      accept(c, mov, register_(c, rax, rdx, S8Selection));
+      register_(c, rsp)->accept(c, add, immediate(c, 16));
+      accept(c, mov, register_(c, rax, rdx, Select8));
     }
   } break;
 
   case mov: {
-    if (selection == DefaultSelection) {
-      encode(c, 0x89, operand->value(c), this, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      encode(c, 0x89, operand->value(c), this, operand->selection);
     } else {
-      switch (selection) {
-      case S1Selection:
-        if (operand->value(c) > rbx) {
-          c->code.append(0x40);
+      switch (operand->selection) {
+      case Select1:
+        if (BytesPerWord == 8) {
+          if (operand->value(c) > rbx) {
+            c->code.append(0x40);
+          }
+          encode(c, 0x88, operand->value(c), this, false);
+        } else {
+          if (operand->value(c) > rbx) {
+            RegisterOperand* ax = temporary(c, rax);
+            ax->accept(c, mov, register_(c, operand->value(c)));
+            accept(c, mov, register_(c, rax, NoRegister, Select1));
+            ax->release(c);
+          } else {
+            encode(c, 0x88, operand->value(c), this, false);
+          }
         }
-        encode(c, 0x88, operand->value(c), this, false);
         break;
 
-      case S2Selection:
-      case Z2Selection:
+      case Select2:
+      case Select2z:
         c->code.append(0x66);
         encode(c, 0x89, operand->value(c), this, false);
         break;
 
-      case S4Selection:
+      case Select4:
         encode(c, 0x89, operand->value(c), this, false);
         break;
 
-      case S8Selection:
-        assert(c, operand->selection == S8Selection);
-
+      case Select8:
         memory(c, base, displacement, index, scale)->accept
           (c, mov, register_(c, operand->value(c)));
 
@@ -1804,7 +1973,9 @@ MemoryOperand::accept(Context* c, Operation operation,
   } break;
 
   case mul: {
-    if (selection == DefaultSelection) {
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
       RegisterOperand* tmp = temporary(c);
 
       tmp->accept(c, mov, this);
@@ -1813,6 +1984,8 @@ MemoryOperand::accept(Context* c, Operation operation,
     
       tmp->release(c);
     } else {
+      assert(c, operand->selection == Select8);
+
       RegisterOperand* tmp = temporary(c);
       RegisterOperand* ax = temporary(c, rax);
       RegisterOperand* dx = temporary(c, rdx);
@@ -1843,11 +2016,19 @@ MemoryOperand::accept(Context* c, Operation operation,
   } break;
 
   case or_: {
-    encode(c, 0x09, operand->value(c), this, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      encode(c, 0x09, operand->value(c), this, operand->selection);
+    } else {
+      abort(c);
+    }
   } break;
 
   case rem: {
-    if (selection == DefaultSelection) {
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
       RegisterOperand* ax = temporary(c, rax);
       RegisterOperand* dx = temporary(c, rdx);
       ax->accept(c, mov, this);
@@ -1863,41 +2044,62 @@ MemoryOperand::accept(Context* c, Operation operation,
       ax->release(c);
       dx->release(c);
     } else {
-      assert(c, selection == S8Selection);
+      assert(c, operand->selection == Select8);
 
       operand->apply(c, push);
-      apply(c, push);
+      memory(c, base, displacement, index, scale, Select8)->apply(c, push);
       immediate(c, reinterpret_cast<intptr_t>(moduloLong))->apply(c, call);
-      accept(c, mov, register_(c, rax, rdx, S8Selection));
+      register_(c, rsp)->accept(c, add, immediate(c, 16));
+      accept(c, mov, register_(c, rax, rdx, Select8));
     }
   } break;
 
   case shl: {
-    RegisterOperand* cx = temporary(c, rcx);
-    cx->accept(c, mov, operand);
-    encode(c, 0xd3, 4, this, true);
-    cx->release(c);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      RegisterOperand* cx = temporary(c, rcx);
+      cx->accept(c, mov, operand);
+      encode(c, 0xd3, 4, this, operand->selection);
+      cx->release(c);
+    } else {
+      abort(c);
+    }
   } break;
 
   case shr: {
-    RegisterOperand* cx = temporary(c, rcx);
-    cx->accept(c, mov, operand);
-    encode(c, 0xd3, 5, this, true);
-    cx->release(c);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      RegisterOperand* cx = temporary(c, rcx);
+      cx->accept(c, mov, operand);
+      encode(c, 0xd3, 5, this, operand->selection);
+      cx->release(c);
+    } else {
+      abort(c);
+    }
   } break;
 
   case ushr: {
-    RegisterOperand* cx = temporary(c, rcx);
-    cx->accept(c, mov, operand);
-    encode(c, 0xd3, 7, this, true);
-    cx->release(c);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      RegisterOperand* cx = temporary(c, rcx);
+      cx->accept(c, mov, operand);
+      encode(c, 0xd3, 7, this, operand->selection);
+      cx->release(c);
+    } else {
+      abort(c);
+    }
   } break;
 
   case sub: {
-    if (selection == DefaultSelection) {
-      encode(c, 0x29, operand->value(c), this, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      encode(c, 0x29, operand->value(c), this, operand->selection);
     } else {
-      assert(c, selection == S8Selection);
+      assert(c, operand->selection == Select8);
 
       RegisterOperand* ax = temporary(c, rax);
       RegisterOperand* dx = temporary(c, rdx);
@@ -1915,15 +2117,23 @@ MemoryOperand::accept(Context* c, Operation operation,
   } break;
 
   case subb:
-    if (operand->selection == DefaultSelection) {
-      encode(c, 0x19, operand->value(c), this, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      encode(c, 0x19, operand->value(c), this, operand->selection);
     } else {
       abort(c);
     }
     break;
 
   case xor_: {
-    encode(c, 0x31, operand->value(c), this, true);
+    if (operand->selection == DefaultSelection
+        or operand->selection == Select4)
+    {
+      encode(c, 0x31, operand->value(c), this, operand->selection);
+    } else {
+      abort(c);
+    }
   } break;
 
   default: abort(c);
@@ -2111,7 +2321,7 @@ class MyCompiler: public Compiler {
     MyStack* stack = static_cast<MyStack*>(s);
     while (count) {
       -- count;
-      stack = ::pushed(&c, stack, BytesPerWord);
+      stack = ::pushed(&c, stack);
     }
     return stack;
   }
@@ -2122,17 +2332,10 @@ class MyCompiler: public Compiler {
 
   virtual Operand* stack(Stack* s, unsigned index) {
     MyStack* stack = static_cast<MyStack*>(s);
-    unsigned i = 0;
 
-    if (stack->value->footprint(&c) / BytesPerWord == 2) {
-      ++ i;
-    }
-
-    for (; i < index; ++i) {
+    while (index) {
+      -- index;
       stack = stack->next;
-      if (stack->value->footprint(&c) / BytesPerWord == 2) {
-        ++ i;
-      }
     }
 
     return stack->value;
@@ -2163,7 +2366,7 @@ class MyCompiler: public Compiler {
   }
 
   virtual Operand* temporary() {
-    RegisterOperand* r = register_(&c, NoRegister);
+    RegisterOperand* r = register_(&c);
     appendAcquire(&c, r);
     return r;
   }
@@ -2219,7 +2422,9 @@ class MyCompiler: public Compiler {
   }
 
   virtual Operand* result() {
-    return ::temporary(&c, rax, rdx);
+    RegisterOperand* r = register_(&c, rax, rdx);
+    appendAcquire(&c, r);
+    return r;
   }
 
   virtual void return_(Operand* v) {
@@ -2339,6 +2544,10 @@ class MyCompiler: public Compiler {
     return static_cast<MyOperand*>(v)->select(&c, Select2);
   }
 
+  virtual Operand* select2z(Operand* v) {
+    return static_cast<MyOperand*>(v)->select(&c, Select2z);
+  }
+
   virtual Operand* select4(Operand* v) {
     return static_cast<MyOperand*>(v)->select(&c, Select4);
   }
@@ -2347,20 +2556,8 @@ class MyCompiler: public Compiler {
     return static_cast<MyOperand*>(v)->select(&c, Select8);
   }
 
-  virtual Operand* signExtend1(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, SignExtend1);
-  }
-
-  virtual Operand* signExtend2(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, SignExtend2);
-  }
-
-  virtual Operand* zeroExtend2(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, ZeroExtend2);
-  }
-
-  virtual Operand* signExtend4(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, SignExtend4);
+  virtual Operand* select8From4(Operand* v) {
+    return static_cast<MyOperand*>(v)->select(&c, Select8From4);
   }
 
   virtual void prologue() {
diff --git a/src/compiler.h b/src/compiler.h
index 28b585f11c..5dce7d484a 100644
--- a/src/compiler.h
+++ b/src/compiler.h
@@ -36,13 +36,10 @@ class Compiler {
 
   virtual Operand* select1(Operand*) = 0;
   virtual Operand* select2(Operand*) = 0;
+  virtual Operand* select2z(Operand*) = 0;
   virtual Operand* select4(Operand*) = 0;
   virtual Operand* select8(Operand*) = 0;
-
-  virtual Operand* signExtend1(Operand*) = 0;
-  virtual Operand* signExtend2(Operand*) = 0;
-  virtual Operand* zeroExtend2(Operand*) = 0;
-  virtual Operand* signExtend4(Operand*) = 0;
+  virtual Operand* select8From4(Operand*) = 0;
 
   virtual Operand* stack() = 0;
   virtual Operand* base() = 0;